AI NewsWaveOct. 24 AI Briefing: PHP Vulnerability and AI Attacks, Kimi K3.1 Pro Released
Key points
- AI changes the vulnerability race, with response speed at issue
- Performance ratings and overseas expansion of Chinese AI models
- Google Cloud and Meta expand agent services
Coverage continued on a critical security vulnerability in PHP. Boannews reported an analysis saying a North Korea-linked hacking group used AI to prepare attacks just one day after a patch was released. Digital Daily reported that the PHP Foundation admitted its delayed response was a "structural failure," saying it found clues to the problem in 2024 but the fix was late because of compatibility concerns, a shortage of specialists and consensus-based decision-making. Chosun Biz wrote that concern over a "bugmageddon" is growing as AI uncovers even long-undetected vulnerabilities.
China's Moonshot AI unveiled a new flagship model, Kimi K3.1 Pro. AI Times reported that the model ranked fourth overall and first among open models in the Artificial Analysis performance evaluation (AAII). Digital Today, citing SiliconANGLE, wrote that it outperformed Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on some tasks. GeekNews described it as a model in the 3-trillion-parameter class supporting a 1 million-token context. The way performance is compared varies by evaluation method and by outlet.
Google Cloud unveiled Agent Compute Engine, which gives each AI agent its own isolated virtual execution environment. AI Times reported the news on the 24th and said it isolates security with microVMs. AI Times also reported the same day that Anthropic launched a service that lets users use all features of the Claude platform on Google Cloud.
Meta's AI assistant Muse has launched in Japan. According to AI Times, Meta opened the Japan service on the 23rd, its first overseas expansion a little over a month after the North American release in September. Naver said it would review whether to allow it if it is introduced in Korea. Digital Today, citing Japan Today, reported that Japan is the first launch country outside North America and that users can give instructions by text or voice in a dedicated app.
China has allowed DeepSeek and MiniMax to release their models overseas. AI Times reported that the two companies completed the export registration process for releasing open-source models abroad and offering licenses. IT Chosun described the same measure as official approval for overseas use, and analyzed it as a two-track control that blocks the leak of core technology, contrasting it with the recent blocking of Meta's U.S. acquisition of Manus.
The target scale of the government's frontier AI project came under scrutiny. E-Focus reported that the target set by Deputy Prime Minister Bae Kyung-hoon ranges from 2 trillion to 10 trillion parameters, and that the lower bound is smaller than a model whose weights Moonshot AI released three months ago. At the National Assembly audit on the 23rd, Deputy Prime Minister Bae said the project is not being pursued on the president's instruction. iNews24 described it as a 16 trillion won project, while Digital Today described it as a 4.7 trillion won project.
According to AI Times, Anthropic succeeded in mathematically verifying the security properties of 123 open-source software libraries using Claude. In a matter of weeks, it proved that for 239 test entry points, the program does not crash under certain conditions no matter what input is given, and it also found bugs that existing tests had missed.